A deep learning algorithm to translate and classify ... · 9/28/2020 · We showed that a deep...

A deep learning algorithm to translate and classify cardiac

electrophysiology: From iPSC-CMs to adult cardiac cells

Parya Aghasafari PhD1, Pei-Chi Yang PhD1, Divya C. Kernik PhD2, Kauho Sakamoto PhD4, Yasunari

Kanda PhD5, Junko Kurokawa PhD4, Igor Vorobyov PhD1,6 and Colleen E. Clancy PhD1,3

1 Department of Physiology and Membrane Biology

University of California Davis, Davis, CA

2 Washington University in St. Louis

4Department of Bio-Informational Pharmacology

School of Pharmaceutical Sciences

University of Shizuoka

Shizuoka, Japan

5Division of Pharmacology

National Institute of Health Sciences

Kanagawa, Japan

6Department of Pharmacology

University of California, Davis, CA

3 Correspondence: Colleen E. Clancy, Ph.D.

University of California, Davis

Tupper Hall, RM 4303

Davis, CA 95616-8636

Email: [email protected]

Phone: 530-754-0254

Word count:

Subject code: Artificial Intelligence, Machine Learning, Deep Learning, Computational Biology,

Arrhythmias, Pharmacology.

Short title: Translation of cardiac action potential from immature to mature

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted September 29, 2020. ; https://doi.org/10.1101/2020.09.28.317461doi: bioRxiv preprint

mailto:[email protected]

https://doi.org/10.1101/2020.09.28.317461

Abstract

Exciting developments in both in vitro and in silico technologies have led to new ways to identify

patient specific cardiac mechanisms. The development of induced pluripotent stem cell-derived

cardiomyocytes (iPSC-CMs) has been a critical in vitro advance in the study of patient-specific

physiology, pathophysiology and response to drugs. However, the iPSC-CM methodology is limited

by the low throughput and high variability of resulting electrophysiological measurements.

Moreover, the iPSC-CMs generate immature action potentials, and it is not clear if observations in

the iPSC-CM model system can be confidently interpreted to reflect impact in human adults.

There has been no demonstrated method to allow reliable translation of results from the iPSC-CM

to a mature adult cardiac response. Here, we demonstrate a new computational approach

intended to address the current shortcomings of the iPSC-CM platform by developing and

deploying a multitask network that was trained and tested using simulated data and then applied

to experimental data. We showed that a deep learning network can be applied to classify cells

into the drugged and drug free categories and can be used to predict the impact of

electrophysiological perturbation across the continuum of aging from the immature iPSC-CM

action potential to the adult ventricular myocyte action potential. We validated the output of the

model with experimental data. The method can be applied broadly across a spectrum of aging,

but also to translate data between species.


https://doi.org/10.1101/2020.09.28.317461

Introduction

The development of novel technologies has resulted in new ways to study cardiac function and

rhythm disorders [1]. One such technology is the induced pluripotent stem cell-derived

cardiomyocyte (iPSC-CMs) in vitro model system [2]. The iPSC-CM system constitutes a powerful

in vitro tool for preclinical assessment of cardiac electrophysiological impact and drug safety

liabilities in a human physiological context [3-8]. Moreover, because iPSC-CMs can be cultured

from patient specific-cells, it is one of the new model systems for patient-based medicine.

While in vitro iPSC-CM utilization allows for observation of a variety of responses to drugs and

other perturbations [9-12], there is still a major inherent limitation: The complex differentiation

process to create iPSC-CMs results in a model of cardiac electrical behavior, which is relatively

immature, resembling fetal cardiomyocytes. Hallmarks of the immature phenotype include

spontaneous beating, immature calcium handling, presence of developmental currents, and

significant differences in the relative contributions of repolarizing potassium currents compared

to adult ventricular myocyte [13-15]. The profound differences between the immature iPSC-CM

and the adult cardiac myocyte have led to persistent questions about the utility of the iPSC-CM

action potential to predict relevant impact on adult human electrophysiology [16, 17]. In this

study, we set out to take a first step to bridge the gap.

Here we describe a new way to connect data from patient derived iPSC-CMs to feed the

development of computational models and to fuel the application of machine learning techniques

to allow a way to predict, classify and translate changes observed in cardiac activity from the in

vitro iPSC-CMs to predict their effects on adult cardiomyocytes (adult-CMs). The iPSC-CM and

adult-CM action potential (AP) populations were paced with physiological noise with 0-50% IKr

block to generate a robust simulated data set to train and test the deep learning algorithm. We

then showed how it can be applied even to scarce experimental data, which was also used to

validate the model.


https://doi.org/10.1101/2020.09.28.317461

Deep learning techniques are increasingly used to advance personalized medicine [18]. Long-

short-term-memory (LSTM) based networks are capable of learning order dependence in

sequence prediction problems [19] and have been widely used for cardiac monitoring purposes.

They have been used to extract important biomarkers from raw ECG signals [20-22] and help

clinicians to accurately detect common heart failure features in ECG screenings [20, 23-27]. LSTM

classifiers have also been employed to automatically classify arrhythmias using ECG features [28-

32]. In addition, modeling and simulation, and multi-task networks have been widely implemented

to identify hERG blockers from sets of small molecules in drug discovery and screening [33-38].

Here, we developed and applied a multitask network leveraging LSTM architecture to classify and

translate observations from cardiac activity of in vitro iPSC-CMs to predict corresponding adult

ventricular myocyte cardiac activity. Here, we show that developments in iPSC-CM experimental

technology and developments in cardiac electrophysiological modeling and simulation of iPSC-

CMs can be leveraged in a new machine learning model to interrogate disease and drug response

in cardiac myocytes from immaturity to maturation.

Results

In this study, we set out to develop and apply a deep learning multitask network that would

perform two distinct tasks: A) The first task is cellular level action potential classification to

distinguish between drug free action potential waveforms and action potential waveforms

following application of IKr block (drug-induced hERG channel block). B) The second goal is cellular

level action potential translation to translate immature action potential waveforms to predict

adult waveforms. We first simulated a population of 542 in silico iPSC-CMs and O’Hara-Rudy in

silico human adult ventricular action potentials (adult-CMs). An average cell AP from the

population is shown in Figure 1A for iPSC-CMs and Figure 1B for adult-CMs. The ionic currents

underlying the in silico iPSC-CM APs and adult-CMs APs are markedly different as shown in panels

C and D, respectively. In Figure 1E, the population of iPSC-CM APs is shown and the adult-CM AP

population is in Figure 1F, generated by applying physiological noise as we have described

previously [39-41]. In Figure 1 panels G and H, the impact of perturbation of the in silico iPSC-CM


https://doi.org/10.1101/2020.09.28.317461

APs and adult-CMs APs by 1-50% IKr block is shown. Panel I shows the parameter comparison for

iPSC-CMs and adult-CMs APs without and with perturbation by 1-50% IKr block.

Figure 1. Cellular action potential (AP) and ionic currents for induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) and adult-CMs (O’Hara-Rudy human ventricular action


https://doi.org/10.1101/2020.09.28.317461

potentials). A - B Comparison of Cellular APs in the baseline model of iPSC-CMs and control case of adult-CMs at cycle length of 982 ms. C. Simulated ionic currents (ICaL, IKr, IKs, Ito, IK1) during

baseline iPSC-CMs AP compared to the simulated currents profiles during control case of adult-CMs AP in D. E. iPSC-CMs APs of spontaneously beating cells (n = 304) were simulated after incorporating physiological noise. F. Adult-CMs APs were simulated with physiological noise

currents at matching cycle lengths of iPSC-CMs in panel E. G. iPSC-CMs APs were simulated with perturbation by 1-50% IKr block after incorporating physiological noise. H. Adult-CMs APs were

simulated with perturbation by 1-50% IKr block and physiological noise currents at matching cycle lengths of iPSC-CMs in panel G. I. Comparison between iPSC-CMs and adult-CMs for upstroke

velocity, maximum diastolic potential (MDP) and action potential durations (APD).

We next developed a multitask deep learning network (Figure 2A), for the independent translation

and classification tasks to be performed. The goal of the translation task (Figure 2B) is to use an

immature cardiac action potential waveform dataset and convert (translate) these data to predict

the resulting effect on mature cardiac action potential waveforms. Figure 2C shows the

classification task that is intended to be used to classify action potential waveforms into drug free

(iPSC-CMs APs with less than a determined threshold of (threshold%) IKr block) and drugged (iPSC-

CMs APs with greater than or equal to a determined threshold of (threshold%) IKr block) categories.

We combined these two tasks to demonstrate a positive relationship between planned tasks and

empower the multitask network to predict mature cardiac action potential waveforms with better

accuracy simultaneously. Both tasks are performed by a network comprising two long-short term

memory (LSTM) layers (Figure 2D) followed by independent fully connected layers for each task

(Figure 2E). Notably, the LSTM layers are shared for both tasks (please see methods for details on

the LSTM layers). The features extracted by the LSTM layers are then used as input for the

independent fully connected neural network layers with two hidden layers. The outputs from the

network are both the mature cardiac action potential waveform, and the category of action

potential waveform (drug free or drugged categories).

We used normalized and labeled in silico iPSC-CM APs and adult-CMs APs without (0% IKr block)

(Figure 1 E-F) and with perturbation by 1-50% IKr block (Figure1 G-H) as input and output for

training the multitask network (Figure 2A). First, we started training the network considering drug-

free (0% block) and drugged (1-50% IKr) cases for classification task. Then, we tested if training the


https://doi.org/10.1101/2020.09.28.317461

multitask network with cellular action potential waveforms that had been subject to various IKr

blocking conditions (range 1-50% block) could improve the accuracy of the network for

reconstructing adult-CM APs from iPSC-CM APs . The long short-term memory (LSTM) layer

representation is shown as a schematic in Figure 2D. The computation module and gating

mechanism for each training iteration is also depicted in Figure 2D.

Figure 2. The multitask network architecture. A. The general overview of the multitask network presented in this study. B. The translation task to reconstruct adult-CMs APs from corresponding

iPSC-CMs APs. C. The classification task is shown with IKr block threshold. D. The repeating module in the implemented LSTM layers in the presented multitask network. E. The architecture

of the implemented fully connected layers in the presented multitask network.


https://doi.org/10.1101/2020.09.28.317461

We next explored whether the translation and classification tasks in this study are compatible with

each other, which would allow them to be combined into a single multitask network. The benefits

of combining the tasks are to save computation time and also potentially to improve each task

performance. Therefore, we applied the multitask network to different scenarios in this study as

shown in the workflow schematic in Figure 3. The data generated through the various scenarios

are described in the subsequent figures in the study. We first trained the single task networks

independently and compared the performance for the individual translation (purple, left) and

classification (pink, second from left) tasks to the combined multitask network (yellow in middle).

We applied statistical measures for the translation and classification tasks to evaluate task

alignments. Next, we applied a forward and backward digital filter technique (light blue, second

from right) to iPSC-CMs and adult-CMs APs traces and retrained the multitask network to explore

the effect of noise filtering on the multitask network performance. In the last scenario, we

investigated the importance of carrying out time dependency analysis by removing LSTM layers

for both tasks (gray, right) and retrained the network with various threshold for the IKr block

classification task (red block) to determine the optimal threshold for the best performance of the

multitask network.

Figure 3. Work flow schematic describing different scenarios in this study: Exploring the task alignment by comparing the performance for the individual translation (purple, left) and

classification (pink, second from left) tasks to the combined multitask network (yellow in middle);


https://doi.org/10.1101/2020.09.28.317461

Inspecting the influence of applying noise-filtering techniques (light blue, second from right) to the iPSC-CMs and adult-CMs APs on the multitask network performance for the translation and classification tasks; Investigating the importance of considering the existing time dependency within simulated iPSC-CMs and adult-CMs APs for training the multitask network (gray, right); Tracking the performance of the multitask network for the translation and classification tasks

with various threshold for the IKr block (red block) to identify the optimal threshold for classification task.

From the population of 542 computer model generated iPSC-CMs and adult-CM action potentials;

which contained both drug free and drugged cases, we randomly selected 80% of the virtual cell

waveforms containing both drug free and drugged cases to train the multitask network. The

procedure was as follows: The multitask network utilized the iPSC-CM AP training data with

physiological noise as an input to the network. The network was then optimized to best produce

a translation of these data to match model generated adult-CM action potentials that were

produced under the same pacing conditions and degree of IKr block (i.e. the immature and adult

AP traces were generated through model simulations under the same pacing frequency conditions

and extent of IKr reduction for both). These are shown as the “train” data set in Figure 4 A (iPSC-

CMs) and B (adult-CMs training data in red). The optimized network was able to accurately

translate the training data by generating a similar matched set of adult-CM action potentials as

shown in blue in B (adult-CMs translated from iPSC-CMs by the network). The histogram in Figure

4C shows the good agreement between the model generated adult action potentials and iPSC-CM

APs translated to adult-CM APs in terms of the frequency of virtual cells with similar APD.

The remaining 20% of virtual cell waveforms for both immature and mature APs were designated

as the test set (data “unseen” to the network during training) and used to evaluate the

performance of the multitask network as shown in Figure 4D (iPSC-CMs) and 4E for adult-CMs in

red with superimposed translated data in blue. As indicated by the near superimposition of the

histogram distribution in Figure 4 panel F, the translation from iPSC-CMs to adult-CMs AP across

the range of pacing frequencies was successful. Since drug free and drugged data sets are mixed

in the training and test set, we also plotted the simulated iPSC-CMs AP (Black) and adult-CMs AP

(red) for drug free (Figure 4 G-H) and drugged AP waveforms (Figure 4J-K) separately to clarify the

performance of the multitask network for translating iPSC-CMs to adult-CMs AP for categories


https://doi.org/10.1101/2020.09.28.317461

considered in the classification task. Superimposed translated adult-CMs AP waveforms are also

illustrated in panel H and K in blue for drug free and drugged categories, respectively. Panel I and

L illustrates the superimposition of the histogram distribution of APD90 values for drug free and

drugged categories.

To evaluate the network performance for classification task we used area under receiver operating

characteristic plot [42], recall, precision and F1_score [43] as statistical measures (Table 1). We

compared the values for discussed measures for both training and test set and observed that the

network can categorize APs into drug-free and drugged waveforms with approximately 90%

accuracy (Table 1).

Figure 4. The performance of multitask network for translating iPSC-CMs APs into adult-CMs APs. A. The iPSC-CMs APs used for training the multitask network contained a variety of action potential


https://doi.org/10.1101/2020.09.28.317461

morphologies without and with IKr block (Train set). B. Comparison between simulated (red) and translated adult-CM APs (blue) in the training set. C. Comparison between the histogram distribution of APD90 values for simulated and translated adult-CM APs in the training set. D. Dedicated iPSC-CM APs for testing the performance of multitask network for the unseen AP data to the network (Test set) E. Comparison between simulated (red) and translated adult-CM APs (blue) in the test set. F. Comparison between histogram distribution of APD90 values for simulated and translated adult-CM APs in the test set. G. Population of iPSC-CM APs without IKr block. H. Comparison between simulated (red) and translated adult-CM APs (blue) without IKr block. I. Comparison between histogram distribution of APD90 values for simulated and translated adult-CM APs without IKr block. J. Population of iPSC-CMs APs with IKr block. K. Comparison between simulated (red) and translated adult-CM APs (blue) with IKr block. L. Comparison between histogram distribution of APD90 values for simulated and translated adult-CM APs with IKr block.

Table 1. Statistical measures for evaluating the performance of the multitask network for classifying AP traces into ones with and without 1-50% IKr block for both training and test sets.

Measures AUROC Recall Precision F1-score

Training set performance 0.85 0.75 0.96 0.84

Test set performance 0.90 0.85 0.95 0.9

To evaluate the task alignment, we trained two single task networks for classification and

translation tasks separately and compared the single task network performance with multitask

network performance (Table 2). We compared the statistical measures for both translation and

classification tasks. Interestingly, the multitask network actually performed better in drugged

cases (Figure 4L) compared to drug-free cases (Figure 4I). As it can be seen in the presented

measures, the multitask network has less mean-squared-error (MSE) and APD90 calculation error

for translating the traces with 1-50% IKr block (Table 2). The implication is that addition of the

classification task will improve the reliability of the network to translate iPSC-CMs APs into adult-

CMs AP.

In contrast to the benefit of the multitask network to improving translation accuracy, no significant

improvement was observed in statistical measures for classification task (Table 2). Notably, the

statistical measures are close in value for the single task and multitask network which indicates


https://doi.org/10.1101/2020.09.28.317461

good task alignment in the multitask network. Therefore, combining the two tasks may indeed

save computation time and also improve the translation task performance (Table2).

Table 2. Statistical measures for evaluating the performance of the multitask network for both translation and classification tasks to explore task alignment.

Translation

Networks MSE R2_score APD90 error for block

Single task (Translation) 0.0014 0.995 3.40%

Single task (Classification) - - -

Multitask 0.0013 0.995 3.25%

Classification

Networks AUROC Recall Precision F1-score

Single task (Translation) - - - -

Single task (Classification) 0.90 0.85 0.95 0.9

Multitask 0.90 0.85 0.95 0.9

Next, in order to demonstrate how the multitask network can be extended to apply even to noisy

experimental data, we applied a digital filter forward-backward technique [44] to iPSC-CM and

adult-CM AP traces without and with perturbation by 1-50% IKr block (Figure 5). Panel A shows

simulated drug-free iPSC-CM with physiological noise in green and after applying a noise filtering

technique in purple. To assess the phase distortion resulting from noise filtering for APD90 values,

we then plotted a histogram distribution of APD90 values in Figure 5B, which shows near

superimposition of the histogram distribution. This indicates that noise filtering does not change

AP waveforms shape and only removes existing vertical noise. Panels C and D show the same

process as described in panels A and B, but for the adult-CM APs (again, with physiological noise

in green and after applying noise filtering techniques in purple in panel C, and histogram

distributions of APD90 in panel D) .

In Figure 5E and F, we show the drugged AP waveforms with physiological noise in green, noise-

filtered in purple and without physiological noise in pink for iPSC-CMs and corresponding

histograms indicating the frequency of APD90 values, respectively. The adult-CM (Figure 5G) is

shown for drugged AP waveforms with physiological noise in green, noise-filtered in purple and


https://doi.org/10.1101/2020.09.28.317461

without physiological noise in pink along with the overlaid APD90 histograms in Figure 5H. We also

undertook retraining of the multitask network with noise-filtered traces to explore the effect of

noise filtering on the network performance. We compared the statistical measures for both

translation and classification tasks for the trained network based on AP waveforms with

physiological noise and noise-filtered traces (Table 3) and observed that applying noise filtering

techniques could slightly improve multitask network performance for both tasks.

Table 3. Statistical measures for evaluating the influence of applying noise-filtering techniques and considering the existing time dependency within simulated iPSC-CMs and adult-CMs APs on the multitask network performance.

Translation

Networks MSE R2_score APD90 error with Ikr block

Multitask without noise-filtering 0.0013 0.995 3.25%

Multitask with noise-filtering 0.0013 0.995 3.17%

Multitask with noise-filtering and removing LSTM layers

0.0016 0.980 3.6%

Classification

Networks AUROC Recall Precision F1-score

Multitask without noise-filtering 0.90 0.85 0.95 0.90

Multitask with noise-filtering 0.91 0.87 0.95 0.91

Multitask with noise-filtering and removing LSTM layers

0.88 0.85 0.87 0.88


https://doi.org/10.1101/2020.09.28.317461

Figure 5. Applying a digital filter forward and backward technique to iPSC-CM and adult-CM APs without and with IKr block indicates zero phase distortion for APD90 values. A. Simulated drug-

free iPSC-CMs APs with physiological noise in green and after applying the noise filtering technique in purple. B. Comparison between histogram distributions of APD90 values for noisy

and noise-filtered iPSC-CMs APs. C. Simulated drug-free adult-CMs APs with physiological noise in green and after applying noise filtering technique in purple. D. Comparison between

histogram distribution of APD90 values for noisy and noise-filtered adult-CMs APs. E. Simulated iPSC-CM APs with IKr block with physiological noise in green, after applying noise filtering

technique on simulated traces in purple, and without physiological noise in pink. F. Comparison


https://doi.org/10.1101/2020.09.28.317461

between histogram distribution of APD90 values for noisy and noise-filtered versus iPSC-CM APs without physiological noise. G. Simulated adult-CM APs with IKr block with physiological noise in

green, after applying noise filtering technique on simulated traces in purple, and without physiological noise in pink. H. Comparison between histogram distribution of APD90 values for

noisy and noise-filtered versus simulated adult-CMs APs without physiological noise.

We next explored the importance of the existing time dependency within the simulated iPSC-CMs

and adult-CMs APs on the multitask network performance. We did this by extracting the LSTM

layers and compared the network performance with and without the LSTM layers for both

translation and classification tasks (Table 3). As can be seen, removing LSTM layers worsened the

network performance for both tasks. In summary, we determined that training the multitask

network both with noise filtered traces and taking existing time dependence into consideration as

the best approach to design and perform our desired translation and classification tasks.

To test whether training the multitask network with cellular action potential waveforms that had

been subject to a variety of IKr blocking conditions (range 1-50% block) could improve the

performance of the network, we retrained the multitask network with various IKr block

percentages for the classification task and compared the performance of the multitask network

to identify the optimal threshold for the classification task (Figure 6). For our test set, we recorded

the MSE (Figure 6A) and the error in calculation of APD90 (Figure 6B) to investigate its influence on

the translation task and monitored values of AUROC (Figure 6C) and precision (Figure 6D) for the

classification task. As can be seen in Figure 6A-D, perturbation by 25% IKr block resulted in the

best performance for both the translation and classification tasks when the selected threshold

defined two classes: class 1 (iPSC-CMs APs with less than 25% IKr block) and class 2 (iPSC-CMs APs

with greater than or equal to 25% IKr block). In summary, the multitask network with noised-

filtered iPSC-CMs APs and adult-CMs APs with LSTM layers that defined a 25% IKr threshold for the

classification task emerged as the best proposed network.


https://doi.org/10.1101/2020.09.28.317461

Figure 6. Tracking the performance of the multitask network for translation and classification tasks with different IKr block percentage thresholds to identify an optimal one. A. Calculated the

mean-squared-error (MSE) between simulated and translated adult-CM AP waveforms by changing IKr block threshold for the classification task for test set. B. Calculated error between

APD90 values for simulated and translated adult-CM APs for test set. C. Multitask network AUCROC measure for classifying iPSC-CM APs with IKr block into class 1 (iPSC-CM APs with less

than threshold % IKr block) and class 2 (iPSC-CM APs with greater than or equal to threshold % IKr block). D. Multitask network precision measure for classifying iPSC-CM APs with IKr block into

class 1 and 2.

We next set out to demonstrate the real-world utility of the multitask classification and translation

network by applying the network to experimental data. We utilized experimental iPSC-CM APs

from the Kurokawa lab (Figure 7A) and applied the translation task resulting in the predicted adult-

CM APs as shown in Figure 7B. The translation notably resulted in a reduction in variability in APD

in the adult translated cells, consistent with our simulated results and with previous experimental

observations [16, 45]. In an additional validation of the deep learning network, we first simulated

iPSC-CM APs with 50% block of IKr. We then used these simulation APs as an input for the multitask

network and utilized the output from the translation task as a prediction of the effect of 50% block

on adult-CMs. In Figure 7C, the translated drugged APD90 values are shown as turquoise asterisks


https://doi.org/10.1101/2020.09.28.317461

plotted with directly simulated O’Hara-Rudy adult-CMs APs with 50% IKr block (red curve) and

additional validation by experimental 50% block of IKr by 1M E-4031 (blue squares) [40]. These

data strongly suggest that the effects of drug block in iPSC-CMs can successfully be translated to

predict its effect on adult human APs.

Figure 7. Translation of experimentally recorded iPSC-CMs APs into adult-CMs APs to validate the multitask network performance. A. Experimentally recorded iPSC-CMs APs from the Kurokawa lab. B. Translated adult-CMs APs from experimentally recorded iPSC-CMs APs using presented

multitask network. C. Comparing translated adult-CMs AP APD90 values with 50% IKr block (turquoise asterisks) with previously published simulated (red curve for drugged and black for drug-free control) and experimental (blue squares) values in O’Hara-Rudy study [40] indicates

successful model validation.

Methods

Generation of the in silico data for training and testing the deep learning model:

iPSC-CM and adult-CM baseline Action potentials with or without IKr block


https://doi.org/10.1101/2020.09.28.317461

iPSC-CMs baseline cells were paced from steady-state. The model formulation used for the

baseline adult-CMs was the O’Hara-Rudy endocardial model cell [40]. The control adult-CMs was

paced at cycle length of 982 ms to match the cycle length of the last beat of iPSC-CMs AP. The

block case of adult-CMs (50% inhibition) was paced at cycle length of 1047 ms to match the cycle

length of the last beat of iPSC-CMs AP with 50% IKr block.

iPSC-CMs and adult-CMs baseline simulations with physiological noise currents

Data set 1: Without IKr block: iPSC-CMs AP populations (n = 304) were generated after

incorporating physiological noise. The adult-CMs were paced with noise for 100 beats after

reaching steady state at matching cycle length of the last beat of iPSC-CMs AP populations.

Data set 2: With IKr block: The adult-CMs with noise were paced with 1-50% IKr block. The model

was simulated at five varying beating rates for each percentage of block that match to the last

beat of iPSC-CMs with 1-50% IKr block (n = 250). The numerical method used for updating the

voltage was Forward Euler method [46].

Simulated physiological noise currents:

Simulated noise current was added to the last 100 paced beats in the simulated AP models, and

simulated action potentials (APs) were recorded at the 2000th paced beat in single cells. This

noise current was modeled using the equation from [39],

𝑉𝑡+∆𝑡 = 𝑉𝑡 −𝐼(𝑉𝑡)∆𝑡

𝐶𝑚+ 𝜉𝑛√∆𝑡 (1)

Where n is N(0,1) is a random number from a Gaussian distribution, and ∆t is the time step. =

0.3 is the diffusion coefficient, which is the amplitude of noise. The noise current was generated

and applied to membrane potential Vt throughout the last 100 beats of simulated time course.

Experimental iPSC-CMs:

Human iPSC-CMs (201B7, RIKEN BRC, Tsukuba, Japan) were cultured and subcultured on SNL76/7

feeder cells as described in detail previously [47]. Cardiomyocyte differentiation was performed

either by the EB method with slight modifications [47]. Commercially available iCell-

cardiomyocytes (FUJIFILM Cellular Dynamics, Inc., Tokyo, Japan) were cultured according to the


https://doi.org/10.1101/2020.09.28.317461

manual provided from the companies. Action potentials were recorded with the perforated

configuration of the patch-clamp technique as described in detail previously [47]. Measurements

were performed at 36 ± 1 °C with the external solution composed of (in mM): NaCl (135), NaH2PO4

(0.33), KCl (5.4), CaCl2 (1.8), MgCl2 (0.53), glucose (5.5), HEPES, pH 7.4. To achieve patch

perforation (10-20 MΩ; series resistances), amphotericin B (0.3-0.6 µg/mL) was added to the

internal solution composed of (in mM): aspartic acid (110), KCl (30), CaCl2 (1), adenosine-5’-

triphosphate magnesium salt (5), creatine phosphate disodium salt (5), HEPES (5), EGTA (11), pH

7.25. In quiescent cardiomyocytes, action potentials were elicited by passing depolarizing current

pulses (2 ms in duration) of suprathreshold intensity (120 % of the minimum input to elicit action

potentials) with a frequency at 1 Hz unless noted otherwise.

The multitask network architecture:

The proposed multitask network comprises two LSTM layers followed by independent fully

connected layers (Figure 2A) for the classification and translation tasks. The LSTM layers

memorizes the important time dependent information and then transfers the extracted

information (features) into the subsequent fully connected layers to translate immature cardiac

action potential waveforms into mature cardiac action potential waveforms (Figure 2B) and

classify iPSC-CM APs into class 1 (iPSC-CM APs with less than the threshold % IKr block) and class 2

(iPSC-CM APs with greater than or equal to the threshold % IKr block) (Figure 2C).

Long-short term memory (LSTM) layers (Figure 2D):

We used LSTMs as the first two layers of the multi-task network to promote network learning for

which data in a sequence is important to keep or to throw away. At each time step, the LSTM cell

takes in three different pieces of information, the current input data (𝐴𝑃𝑖𝑃𝑆𝐶𝑡), the short-term

memory (hidden state) from the previous cell (ℎ𝑡−1) and the long-term memory (cell state)

(𝐶𝑡−1). The LSTM cells contain internal mechanisms called gates. The gates are neural networks

(with weights (w) and bias terms (b)) that regulate the flow of information at each time step before

passing on the long-term and short-term information to the next cell [48]. These gates are called

the input gate, the forget gate, and the output gate (Figure 2D).


https://doi.org/10.1101/2020.09.28.317461

The forget gate, as the name implies, determines which information from the long-term memory

should be kept or discarded. This is done by multiplying the incoming long-term memory by a

forget vector generated by the current input and incoming short-term memory. To obtain the

forget vector, the short-term memory and current input are passed through a sigmoid function

(σ) [49]. The output vector of sigmoid function is binary comprising 0s and 1s and is then multiplied

by the long-term memory to choose which parts of the long-term memory to retain (Eq. 2).

𝐹𝑡 = 𝜎𝑓(𝑤𝑓𝐴𝑃𝑖𝑃𝑆𝐶𝑡+ 𝑤𝑓ℎ𝑡−1 + 𝑏𝑓) (2)

The input gate decides what new information will be stored in the long-term memory. It considers

the current input and the short-term memory from the previous time step and transforms the

values to be between 0 (unimportant) and 1 (important) using a sigmoid activation function (Eq.

3). The second layer in input gate takes the short-term memory and current input and passes it

through a hyperbolic tangent (tanh) activation function to regulate the network (Eq. 4).

𝐼𝑡 = 𝜎𝑖(𝑊𝑖𝐴𝑃𝑖𝑃𝑆𝐶𝑡+ 𝑊𝑖ℎ𝑡−1 + 𝑏𝑖) (3)

𝑆𝑡 = 𝑡𝑎𝑛ℎ𝑖(𝑤𝑠𝐴𝑃𝑖𝑃𝑆𝐶𝑡+ 𝑤𝑠ℎ𝑡−1 + 𝑏𝑠) (4)

The outputs from the forget and input gates then undergo a pointwise addition to give a new

version of the long-term memory (Eq . 5), which is then passed on to the next cell.

𝐶𝑡 = 𝐹𝑡 ∗ 𝐶𝑡−1 + 𝐼𝑡 ∗ 𝑆𝑡 (5)

Finally, the output gate utilizes current input and previous short-term memory and passes them

into a sigmoid function (Eq. 6). Then the new computed long-term memory passes through a tanh

activation function and the outputs from these two processes are multiplied to produce the new

short-term memory (Eq. 7).

𝑂𝑡 = 𝜎𝑜(𝑤𝑜𝑥𝑡 + 𝑤𝑜ℎ𝑡−1 + 𝑏𝑜) (6)

ℎ𝑡 = 𝑂𝑡 ∗ 𝑡𝑎𝑛ℎ𝑜(𝐶𝑡) (7)

The short-term and long-term memory produced by these gates is carried over to the next cell for

the process to be repeated. The output of each time step is obtained from the short-term memory,


https://doi.org/10.1101/2020.09.28.317461

also known as the hidden state, and is subsequently passed into fully connected layers to perform

the translation and classification tasks.

Fully connected layers (Figure 2E):

The fully connected neural network layers contains input, hidden and output layers (Figure 2E)

which may have various numbers of neurons. Every neuron in a layer is connected to neurons in

another layer [50]. Fully connected layers receive the output of LSTM layers as input. The fully

connected layers calculate a weighted sum of LSTM outputs and add a bias term to the outputs.

These data are then passed to an activation function (f) to define the output for each node (Eq.

8,9) [51].

𝑎𝑗𝑛 = 𝑓(𝑍𝑗

𝑛) (8)

𝑍𝑗𝑛 = 𝑊𝑖,𝑗

𝑛 ∗ 𝑎𝑗𝑛−1 + 𝑏𝑛 (9)

where 𝑎0 is the input data and 𝑎𝑛+1 is output data 𝑦 ̂𝜖 {𝑦𝑡𝑖

, 𝑦𝑐𝑖

} where 𝑦𝑡𝑖 and 𝑦𝑐𝑖

are the outputs for

translation and classification tasks, respectively. We first assign random values to all network

parameters (𝜃𝑡; nodes weight (𝑊𝑖,𝑗), bias term (𝑏) and network hyperparameters and select the

best network infrastructure; number of hidden layers, number of neurons and activation functions

for each hidden layer. Next, we estimate the network errors using mean-squared-error (Eq. 10)

and cross-entropy loss functions (Eq. 11) to map the translation and classification tasks [52, 53],

respectively.

𝑀𝑆𝐸 = 1

𝑛∑ ‖𝑦𝑡𝑖

− �̂�𝑡𝑖‖

2𝑛

𝑖=1 (10)

𝐶𝑟𝑜𝑠𝑠𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = − (𝑦𝑐𝑖log(�̂�𝑐𝑖

) + (1 − 𝑦𝑐𝑖)log (1 − �̂�𝑐𝑖

)) (11)

where n is the total number of input samples and 𝑦𝑡𝑖 and �̂�𝑡𝑖

are the simulated and translated

adult-CM APs (the network output for translation task). The 𝑦𝑐𝑖 is binary indicator of class labels

for iPSC-CM APs (0 for iPSC-CM APs with less than the threshold % IKr block or 1 for iPSC-CM APs

with greater than or equal to the threshold % IKr block) and �̂�𝑐𝑖 is predicted probability of APs being

classified into the discussed classes. We used sum of both loss functions (Eq. 12) to calculate the

overall network error (J) for both translation and classification tasks during the network training

process. We updated network parameters (𝜃𝑡+1) using adaptive momentum estimation (ADAM)


https://doi.org/10.1101/2020.09.28.317461

optimization algorithm [54] based on the average gradient of overall loss function with respect to

the network parameters for 128 randomly selected simulated AP traces (mini-batch=128) at each

training iteration (Eqs. 13-15).

𝐽(𝜃𝑡,𝑖) = 𝐶𝑟𝑜𝑠𝑠𝐸𝑛𝑡𝑟𝑜𝑝𝑦𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛(𝜃𝑡,𝑖) + MSE𝑇𝑟𝑎𝑛𝑠𝑙𝑎𝑡𝑖𝑜𝑛(𝜃𝑡,𝑖) (12)

𝜃𝑡+1,𝑖 = 𝜃𝑡,𝑖 −𝛼 .�̂�𝑡

√�̂�𝑡+𝜖 , 𝜃𝑡,𝑖 𝜖 {𝑊𝑖,𝑗

𝑛 , 𝑏𝑗𝑛} (13)

�̂�𝑡 =𝑚𝑡

1−𝛽1𝑡 , where 𝑚𝑡 = (1 − 𝛽1)𝛻𝐽(𝜃𝑡,𝑖) + 𝛽1𝑚𝑡−1 (14)

�̂�𝑡 =𝜈𝑡

1−𝛽2𝑡 , where 𝜈𝑡 = (1 − 𝛽2) (𝛻𝐽(𝜃𝑡,𝑖))

2+ 𝛽2𝑣𝑡−1 (15)

We used rectified linear unit (ReLu) activation function for hidden layers and dropout

regularization [55] with probability of eliminating any hidden units in fully connected layers equal

to 0.25 for updating model parameters to find global minimum of loss function based on a

predefined learning rate (𝛼) , first and second momentum terms (𝛽1, 𝛽

2 ), and a small term

preventing division by zero (𝜖).

We first started training the network considering drug-free (0% block) and drugged (1-50% IKr)

cases for classification task. Then, we tested a hypothesis whether training the multitask network

with classifying cellular action potential waveforms that had been subject to various IKr blocking

conditions (range 1-50% block) could improve the performance of the network. We used MSE (Eq.

10), R2_score (Eqs. 16-17 below) and the histogram distribution of APD90 as statistical measures to

evaluate the performance of network for translation task and area under receiver operating

characteristic curve (AUROC), recall, precision and F1-score to measure capability of network for

classification task as described below.

We applied a forward and backward digital filter technique[44] into normalized and labeled

simulated iPSC-CM APs and adult-CM APs without and with perturbation by 1-50% IKr block and

utilized them as inputs and outputs in the network. The network architecture is implemented using

Pytorch platform[56]. 80% of the simulated APs were considered for training the network and the

remaining were used to test the performance of the network for an unseen dataset during


https://doi.org/10.1101/2020.09.28.317461

training. The network codes have been made publicly available at Github.

(https://github.com/ClancyLabUCD/Multitask_network)

Statistical measures

As we discussed, we used MSE and cross-entropy as criteria for performance evaluation of

translation and classification tasks. In addition to MSE, we computed R2_score [57] (Eqs. 16,17) to

measure how close the translated adult-CM AP (�̂�𝑡𝑖) are to the simulated adult-CM AP (𝑦𝑡𝑖

). We

compared the histogram distribution of simulated and translated adult-CM APD90 values to

investigate the ability of network to predict them accurately.

�̅�𝑡 = 1

𝑛∑ 𝑦𝑡𝑖

𝑛

𝑖=1 (16)

𝑅2 =∑ (�̂�𝑡𝑖

− �̅�𝑡)𝑖

∑ (𝑦𝑡𝑖 −�̅�𝑡 )𝑖

(17)

We used area under the receiver operating curve (AUROC) to measure the capability of model in

distinguishing between classes [42]. Receiver operating curve (ROC) is a plot of the false positive

rate (FPR, the probability that the network classifies iPSC-CM AP with less than 25% IKr block into

greater than or equal to 25% IKr block) (Eq. 18) versus the true positive rate (TPR) or recall (Eq. 19).

AUROC close to 1 represents an excellent model, which has good measure of separability, while a

poor model has AUROC near 0, which means that it has poor separability.

prediction [58]. In addition, the confusion matrix elements were used to measure statistical

metrics (recall, precision and F1-score) to describe the performance of a classification model [13],

where recall (Eq. 19) is the proportion of actual positives that are correctly identified as such.

Precision measures the proportion of correct positive identifications (Eq. 20) and F1-score is the

harmonic mean of precision and recall (Eq. 21).

FPR =𝐹𝑃

𝐹𝑃+𝑇𝑁 (18)

Recall =𝑇𝑃

𝑇𝑃+𝐹𝑁 (19)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =𝑇𝑃

𝑇𝑃+𝐹𝑃 (20)

F1 = 2 ×Precision × Recall

Precision+ Recall (21)


https://github.com/ClancyLabUCD/Multitask_network

https://doi.org/10.1101/2020.09.28.317461

Discussion

In this study, we have developed a data-driven algorithm intended to address well known

shortcomings in the induced pluripotent stem cell-derived cardiomyocyte (iPSC-CMs) platform. A

known concern with iPSC-CMs is that the data collection results in measurements from immature

action potentials, and it is unclear if these data reliably indicate impact in the adult cardiac

environment. Here, we set out to demonstrate a new deep learning algorithm to allow reliable

translation of results from the iPSC-CM to a mature adult cardiac response. The translation task

yielded superb results with 0.0012 MSE, 0.99 R2_score and less than 3% calculation error in

predicting adult APD90 values from iPSC-CMs inputs.

The multi-task network we present here notably conferred additional benefit over considering the

translation and classification tasks separately. For example, we noted that adding the classification

task to distinguish data from action potentials with and without the application of an IKr blocking

drug could improve the performance of the network. Adding the IKr blocking threshold to the

translation task resulted in up to 7% decrease in MSE value. To pick the best threshold, we tested

a range of thresholds for classification and retrained the network with each of them. The threshold

defined as 25% IKr block led to the highest accuracy for both the translation and classification tasks.

The high values for the classification task statistical measures including AUROC and precision equal

to 0.993 and 0.98, respectively, also indicate the credibility of the multitask network to sort iPSC-

CMs APs into groups with greater than or equal to 25% IKr block from iPSC-CMs APs with less than

25% IKr block.

Importantly, the multitask network presented here performed well even in the setting of the noted

variability in measurements from iPSC-CMs. We utilized a modeling and simulation approach from

our recent study [41] to generate a population of IPSC-CM action potentials that incorporate

variability comparable to that in experimental measurements. Utilizing simulated data presented

a unique opportunity: We were able to generate large amounts of data that were used both to

train and optimize the network and then to test the network with specifically designated distinct


https://doi.org/10.1101/2020.09.28.317461

simulated data sets. Utilizing simulated data to train a deep learning network may constitute a

more widely applicable approach that could be used to train variety of networks to perform

multiple functions where access to comparable experimental data is not feasible.

Following the optimization and demonstration of the network as an accurate tool for both

translating and classifying data, we then used the same network to translate experimentally

obtained data. We showed that the proposed network can effectively take experimental data as

an input from immature iPSC-CMs and translate those data to produce adult action potential

waveforms. It is notable that the variation observed in the adult-CM AP duration is smaller

compared to iPSC-CM AP (Figure 7A-B). This has been observed both experimentally [16, 45] and

in our simulated cell environment [41, 59] . Although the simulated iPSC-CM has a large initial

calcium current (Figure 1B) compared to the simulated adult-CM (Figure 1C), the amplitude of

currents flowing through adult-CM action potential plateau is notably larger. The immature iPSC-

CM cells have low conductance during the AP plateau rendering it comparably higher resistance.

For this reason, small perturbations to the iPSC-CMs have a larger impact on the resulting AP

duration than observed in adult cells [60]. We also used simulated iPSC-CMs subject to 50% block

of IKr. We translated those data to adult-CM APs and then compared the previously reported

impact on adult-CM APs to 50% IKr block from experiments and noted excellent agreement thereby

providing validation using experimental data from adult human cells.

In this study, we show that a deep learning network can be applied to classify cells into the drugged

and drug free categories and can be used to predict the impact of electrophysiological

perturbation across the continuum of aging from the immature iPSC-CM action potential to the

adult ventricular myocyte action potential. We translated experimental immature APs into mature

APs using the proposed network and validated the output of some key model simulations with

experimental data. The multitask network in this study was used for translation of iPSC-CMs to

adult APs, but could be readily extended and applied to translate data across species and classify

data from a variety of systems. Also, another extension of the technology presented here is to

predict the impact of naturally occurring mutations and other genetic anomalies [61].


https://doi.org/10.1101/2020.09.28.317461

References

1. Shaheen, N., et al., Human induced pluripotent stem cell-derived cardiac cell sheets

expressing genetically encoded voltage indicator for pharmacological and arrhythmia

studies. Stem cell reports, 2018. 10(6): p. 1879-1894.

2. Leyton-Mange, J.S., et al., Rapid cellular phenotyping of human pluripotent stem cell-

derived cardiomyocytes using a genetically encoded fluorescent voltage sensor. Stem cell

reports, 2014. 2(2): p. 163-170.

3. Sun, N., et al., Patient-specific induced pluripotent stem cells as a model for familial

dilated cardiomyopathy. Sci Transl Med, 2012. 4(130): p. 130ra47.

4. Lan, F., et al., Abnormal calcium handling properties underlie familial hypertrophic

cardiomyopathy pathology in patient-specific induced pluripotent stem cells. Cell Stem

Cell, 2013. 12(1): p. 101-13.

5. Burridge, P.W., et al., Human induced pluripotent stem cell-derived cardiomyocytes

recapitulate the predilection of breast cancer patients to doxorubicin-induced

cardiotoxicity. Nat Med, 2016. 22(5): p. 547-56.

6. Doss, M.X. and A. Sachinidis, Current challenges of iPSC-based disease modeling and

therapeutic implications. Cells, 2019. 8(5): p. 403.

7. Collins, T.A., M.G. Rolf, and A. Pointon, Current and future approaches to nonclinical

cardiovascular safety assessment. Drug Discovery Today, 2020.

8. Wu, J.C., et al., Towards precision medicine with human iPSCs for cardiac

channelopathies. Circulation research, 2019. 125(6): p. 653-658.

9. Tveito, A., et al., Inversion and computational maturation of drug response using human

stem cell derived cardiomyocytes in microphysiological systems. Scientific reports, 2018.

8(1): p. 1-14.

10. Tveito, A., et al., Computational translation of drug effects from animal experiments to

human ventricular myocytes. Scientific Reports, 2020. 10(1): p. 1-11.

11. Sube, R. and E.A. Ertel, Cardiomyocytes Derived from Human Induced Pluripotent Stem

Cells: An In-Vitro Model to Predict Cardiac Effects of Drugs. Journal of Biomedical

Science and Engineering, 2017. 10(11): p. 527.

12. Navarrete, E.G., et al., Screening drug-induced arrhythmia using human induced

pluripotent stem cell–derived cardiomyocytes and low-impedance microelectrode arrays.

Circulation, 2013. 128(11_suppl_1): p. S3-S13.

13. Lieu, D.K., et al., Mechanism-based facilitated maturation of human pluripotent stem

cell-derived cardiomyocytes. Circ Arrhythm Electrophysiol, 2013. 6(1): p. 191-201.

14. Veerman, C.C., et al., Immaturity of human stem-cell-derived cardiomyocytes in culture:

fatal flaw or soluble problem? Stem Cells Dev, 2015. 24(9): p. 1035-52.

15. Tu, C., B.S. Chao, and J.C. Wu, Strategies for Improving the Maturity of Human Induced

Pluripotent Stem Cell-Derived Cardiomyocytes. Circ Res, 2018. 123(5): p. 512-514.

16. Blinova, K., et al., International multisite study of human-induced pluripotent stem cell-

derived cardiomyocytes for drug proarrhythmic potential assessment. Cell reports, 2018.

24(13): p. 3582-3592.

17. Sala, L., M. Bellin, and C.L. Mummery, Integrating cardiomyocytes from human

pluripotent stem cells in safety pharmacology: has the time come? British journal of

pharmacology, 2017. 174(21): p. 3749-3765.


https://doi.org/10.1101/2020.09.28.317461

18. Alhusseini, M.I., et al., Machine Learning to Classify Intracardiac Electrical Patterns

During Atrial Fibrillation: Machine Learning of Atrial Fibrillation. Circulation:

Arrhythmia and Electrophysiology, 2020. 13(8): p. e008160.

19. Hochreiter, S. and J. Schmidhuber, Long short-term memory. Neural computation, 1997.

9(8): p. 1735-1780.

20. Ballinger, B., et al. DeepHeart: semi-supervised sequence learning for cardiovascular

risk prediction. in Thirty-Second AAAI Conference on Artificial Intelligence. 2018.

21. He, R., et al., Automatic cardiac arrhythmia classification using combination of deep

residual network and bidirectional LSTM. IEEE Access, 2019. 7: p. 102119-102135.

22. Hou, B., et al., LSTM Based Auto-Encoder Model for ECG Arrhythmias Classification.

IEEE Transactions on Instrumentation and Measurement, 2019.

23. Warrick, P. and M.N. Homsi. Cardiac arrhythmia detection from ECG combining

convolutional and long short-term memory networks. in 2017 Computing in Cardiology

(CinC). 2017. IEEE.

24. Oh, S.L., et al., Automated diagnosis of arrhythmia using combination of CNN and LSTM

techniques with variable length heart beats. Computers in biology and medicine, 2018.

102: p. 278-287.

25. Chen, C., et al., Automated arrhythmia classification based on a combination network of

CNN and LSTM. Biomedical Signal Processing and Control, 2020. 57: p. 101819.

26. Wang, L. and X. Zhou, Detection of congestive heart failure based on LSTM-based deep

network via short-term RR intervals. Sensors, 2019. 19(7): p. 1502.

27. Bian, M., et al. An accurate lstm based video heart rate estimation method. in Chinese

Conference on Pattern Recognition and Computer Vision (PRCV). 2019. Springer.

28. Yildirim, O., et al., A new approach for arrhythmia classification using deep coded

features and LSTM networks. Computer methods and programs in biomedicine, 2019.

176: p. 121-133.

29. Wang, E.K., X. Zhang, and L. Pan, Automatic classification of CAD ECG signals with

SDAE and bidirectional long short-term network. IEEE Access, 2019. 7: p. 182873-

182880.

30. Martis, R.J., et al., Application of higher order cumulant features for cardiac health

diagnosis using ECG signals. International journal of neural systems, 2013. 23(04): p.

1350014.

31. Liu, F., et al. A LSTM and CNN based assemble neural network framework for

arrhythmias classification. in ICASSP 2019-2019 IEEE International Conference on

Acoustics, Speech and Signal Processing (ICASSP). 2019. IEEE.

32. Yildirim, Ö., A novel wavelet sequence based on deep bidirectional LSTM network model

for ECG signal classification. Computers in biology and medicine, 2018. 96: p. 189-202.

33. Cai, C., et al., Deep learning-based prediction of drug-induced cardiotoxicity. Journal of

chemical information and modeling, 2019. 59(3): p. 1073-1084.

34. Zhang, Y., et al., Prediction of hERG K+ channel blockage using deep neural networks.

Chemical Biology & Drug Design, 2019. 94(5): p. 1973-1985.

35. Dickson, C.J., C. Velez-Vega, and J.S. Duca, Revealing molecular determinants of hERG

blocker and activator binding. Journal of chemical information and modeling, 2019.

60(1): p. 192-203.

36. Ryu, J.Y., et al., DeepHIT: a deep learning framework for prediction of hERG-induced

cardiotoxicity. Bioinformatics, 2020. 36(10): p. 3049-3055.


https://doi.org/10.1101/2020.09.28.317461

37. Yang, P.-C., et al., A computational pipeline to predict cardiotoxicity: From the atom to

the rhythm. Circulation research, 2020. 126(8): p. 947-964.

38. Li, Z., et al., General principles for the validation of proarrhythmia risk prediction

models: an extension of the CiPA in silico strategy. Clinical Pharmacology &

Therapeutics, 2020. 107(1): p. 102-111.

39. Tanskanen, A.J. and L.H. Alvarez, Voltage noise influences action potential duration in

cardiac myocytes. Mathematical biosciences, 2007. 208(1): p. 125-146.

40. O'Hara, T., et al., Simulation of the undiseased human cardiac ventricular action

potential: model formulation and experimental validation. PLoS computational biology,

2011. 7(5): p. e1002061.

41. Kernik, D.C., et al., A computational model of induced pluripotent stem‐cell derived

cardiomyocytes incorporating experimental variability from multiple data sources. The

Journal of physiology, 2019.

42. Fawcett, T., An introduction to ROC analysis. Pattern recognition letters, 2006. 27(8): p.

861-874.

43. Powers, D.M., Evaluation: from precision, recall and F-measure to ROC, informedness,

markedness and correlation. 2011.

44. Gustafsson, F., Determining the initial states in forward-backward filtering. IEEE

Transactions on signal processing, 1996. 44(4): p. 988-992.

45. Fabbri, A., et al., Required GK1 to suppress automaticity of iPSC-CMs depends strongly

on IK1 model structure. Biophysical Journal, 2019. 117(12): p. 2303-2315.

46. Atkinson, K.E., An introduction to numerical analysis. 2008: John wiley & sons.

47. Li, M., et al., Overexpression of KCNJ2 in induced pluripotent stem cell-derived

cardiomyocytes for the assessment of QT-prolonging drugs. Journal of Pharmacological

Sciences, 2017. 134(2): p. 75-85.

48. Cheng, J., L. Dong, and M. Lapata, Long short-term memory-networks for machine

reading. arXiv preprint arXiv:1601.06733, 2016.

49. Olah, C., Understanding LSTM Networks. Aug. 2015. URL https://colah. github.

io/posts/2015-08-Understanding-LSTMs, 2017.

50. Krogh, A., What are artificial neural networks? Nature biotechnology, 2008. 26(2): p.

195-197.

51. Carugo, O., F. Eisenhaber, and Carugo, Data mining techniques for the life sciences. Vol.

609. 2010: Springer.

52. Goodfellow, I., Y. Bengio, and A. Courville, Deep learning. 2016: MIT press.

53. Murphy, K.P., Machine learning: a probabilistic perspective. 2012: MIT press.

54. Kingma, D.P. and J. Ba, Adam: A method for stochastic optimization. arXiv preprint

arXiv:1412.6980, 2014.

55. Zaremba, W., I. Sutskever, and O. Vinyals, Recurrent neural network regularization.

arXiv preprint arXiv:1409.2329, 2014.

56. Ketkar, N., Introduction to pytorch, in Deep learning with python. 2017, Springer. p.

195-208.

57. Devore, J.L., Probability and Statistics for Engineering and the Sciences. 2011: Cengage

learning.

58. Mason, S.J. and N.E. Graham, Areas beneath the relative operating characteristics

(ROC) and relative operating levels (ROL) curves: Statistical significance and

interpretation. Quarterly Journal of the Royal Meteorological Society: A journal of the


https://colah/

https://doi.org/10.1101/2020.09.28.317461

atmospheric sciences, applied meteorology and physical oceanography, 2002. 128(584):

p. 2145-2166.

59. Kernik, D.C., et al., A computational model of induced pluripotent stem-cell derived

cardiomyocytes for high throughput risk stratification of KCNQ1 genetic variants. PLOS

Computational Biology, 2020. 16(8): p. e1008109.

60. Yang, P.C., et al., A computational modelling approach combined with cellular

electrophysiology data provides insights into the therapeutic benefit of targeting the late

Na+ current. The Journal of physiology, 2015. 593(6): p. 1429-1442.

61. Yoshinaga, D., et al., Phenotype-based high-throughput classification of long QT

syndrome subtypes using human induced pluripotent stem cells. Stem cell reports, 2019.

13(2): p. 394-404.


https://doi.org/10.1101/2020.09.28.317461

A deep learning algorithm to translate and classify ... · 9/28/2020 · We showed that a deep...

Documents

Transcript of A deep learning algorithm to translate and classify ... · 9/28/2020 · We showed that a deep...